In semi-supervised deep learning models, multiple deep neural networks are often utilized for optimization, significantly increasing the complexity of the system. Thus, the demand for lightweight networks has become vital for practical applications. Therefore, we propose a lightweight semi-supervised semantic segmentation method by using pixel-level contrastive learning. First, a lightweight segmentation network is proposed as the base architecture. Here, both standard convolution and atrous convolution are replaced by the depth-wise separable convolution and concentrated comprehensive convolution to reduce the channel number of the encoder and decoder, as well as the network load. Additionally, a residual link structure is incorporated to prevent information loss. Finally, to improve the utilization of available data, we explore a supervised pixel-level contrastive loss function to maximize the limited label information under a semi-supervised context. Comparison experiments on the Pascal VOC 2012 and Cityscapes datasets demonstrate that the proposed method outperforms many state-of-the-art semi-supervised semantic segmentation methods.